← Back to Daily Builds

Daily build note · August 21, 2026

Bot Routine Receipt Desk

Bot Routine Receipt Desk is a local evidence and decision tool for small operators who depend on taught or scheduled automations.

Bot Routine Receipt Desk Ui Polish Build note published Live interactive demo

What shipped

A standalone, offline CLI that verifies whether an automated routine completed each declared step. bot_routine_receipt_desk runs on the Python 3.11+ standard library and creates a deterministic, hash-bound receipt with an ACCEPT / RE-TEACH / RETIRE decision sheet.

Findings use attestable when the supplied files verify them and inferred when the evidence is incomplete. Consequence text comes from a fixed table; the tool generates no prose. Artifact hashes are copied from the run record and are never presented as independently verified. A click-through demo built from actual verifier output is included with the build.

Architecture

  1. **Deterministic and offline** — no network, no model providers; identical

inputs produce byte-identical receipts across independent workspaces.

  1. **Clear evidence labels** — verified findings are separated from suspected

findings, so the receipt does not overstate what the evidence supports.

  1. **Fixed consequence table** — decision-sheet language is looked up, never

generated, keeping the receipt auditable and reproducible.

  1. **Honest artifact hashes** — recorded as claimed by the run record and

labeled as such, never presented as independently re-verified.

  1. **Reuse boundary** — conceptual lineage only (receipt identity discipline

from Agent Receipts Desk, declared-vs-executed comparison from Trace Receipt Bridge); no code copied, imported, invoked, or required — the suite passes with this build copied to an isolated directory.

Trimmed scope

No monitoring, scheduler, daemon, browser or screen access, artifact fetching, model providers, network calls, decision enforcement, operational dashboard, PDF export, or batch mode. A week-long engagement repeats the command for each run. Multi-run analysis is limited to the archive lookup that informs the RETIRE rule.

Limitations

Verifies one run per invocation against supplied JSON only — it cannot discover routines or fetch run records itself; artifact contents are hashed as claimed, not re-fetched; the decision sheet recommends but does not enforce (accept / re-teach / retire remains a human call).

Verification

  • bash verify.sh → **exit 0** (34 unit tests + end-to-end + determinism

double-run byte-compare + overwrite-refusal negative).

  • Scenario A (clean): 0 findings, all steps executed.
  • Scenario B (silent-skip demo, default): step pull-subscriber-stats entirely

absent from the run record while the run claims success → STEP_MISSING_FROM_RUN (skipped_silent), run-level SUMMARY_CONTRADICTS_STEPS, plus ASSERTION_FAILED catching the stale cache-sourced sponsor report. Receipt ID brr-38437913ee7f.

  • Scenario C: one drifted (missing artifact despite ok) vs one

honestly-reported failed_reported, cleanly separated; a skip WITH a stated reason is classified skipped_reported, distinct from silence.

  • Outputs byte-identical across two independent workspaces; tampering, extra

files, and input drift all fail validation; failed regeneration leaves prior outputs untouched with no staging/backup residue.

  • Fixture/output counts: inputs per workspace exactly 2

(inputs/routine.json, inputs/run.json); outputs per verified run exactly 3 (routine-receipt.md, routine-receipt.json, receipt-archive/<run_id>.receipt.json).

Suggested next steps

  1. Record the 90-second side-by-side using scenario B: compare the bot's green

success report with the evidence row showing the skipped step.

  1. DM it to three reachable Grok Bot routine owners with the **$250

paid-upfront routine receipt** offer ($900 three-routine anchor) before any public post. Lead with: "A success message does not prove that every automated step ran."

  1. Capture objections verbatim — especially intake friction ("can I get your

run output into this JSON shape?"). That objection, not price, decides whether the offer survives.

  1. Success = one paid receipt or one owner supplying declared steps + a real

run export.